LLM multilingues

Capability can vary substantially by language, writing system, domain, and task. Supporting a language in an interface does not establish equal quality across languages.

Points clés à retenir

• Measure each important language and task.
• Check tokenization and layout constraints.
• Report language-specific regressions.

Training data coverage affects what a model encounters, while tokenization affects how efficiently text is represented. A passage can require different token counts across languages even when it expresses similar information. This changes practical context limits and serving costs. Cross-lingual transfer can help a model apply patterns learned from one language to another. However, transfer is a capability to measure, not a guarantee that specialized terminology, idioms, or culturally situated questions will be handled correctly. Build an evaluation set for each important language and task. Include natural local examples, mixed-language messages, named entities, and longer documents. Translating an English benchmark alone can introduce unnatural wording or errors that confound the measurement. Review the complete user experience: output language, fonts, text direction, locale formats, citations, and fallback behavior. If the system cannot confidently perform a task in a requested language, communicate that limitation and preserve access to the source. Track regression results by language rather than hiding them in one global average.

1. Imagine 900 test questions in language A with 90% accuracy and 100 in language B with 50% accuracy.
2. The combined score is (810+50)/1000 = 86%, which hides the much weaker result for language B.
3. Report both language-specific results and their sample sizes before deciding where the system is ready to use.

These invented counts illustrate the effect of weighting, not an actual multilingual-model benchmark.